Motivation: Accurate determination of single-nucleotide polymorphisms (SNPs) from next-generation sequencing data is a significant challenge facing bioinformatics researchers. Most current methods use mechanistic models that assume nucleotides aligning to a given reference position are sampled from a binomial distribution. While such methods are sensitive, they are often unable to discriminate errors resulting from misaligned reads, sequencing errors or platform artifacts from true variants.
Results: To enable more accurate SNP calling, we developed an algorithm that uses a trained support vector machine (SVM) to determine variants from .BAM or .SAM formatted alignments of sequence reads. Our SVM-based implementation determines SNPs with significantly greater sensitivity and specificity than alternative platforms, including the UnifiedGenotyper included with the Genome Analysis Toolkit, samtools and FreeBayes. In addition, the quality scores produced by our implementation more accurately reflect the likelihood that a variant is real when compared with those produced by the Genome Analysis Toolkit. While results depend on the model used, the implementation includes tools to easily build new models and refine existing models with additional training data.
Availability: Source code and executables are available from github.com/brendanofallon/SNPSVM/